class SpikeSourcePoissonVertex(
ApplicationVertex, AbstractGeneratesDataSpecification,
AbstractHasAssociatedBinary, AbstractSpikeRecordable,
AbstractProvidesOutgoingPartitionConstraints,
AbstractChangableAfterRun, AbstractReadParametersBeforeSet,
AbstractRewritesDataSpecification, SimplePopulationSettable,
ProvidesKeyToAtomMappingImpl):
""" A Poisson Spike source object
"""
SPIKE_RECORDING_REGION_ID = 0
def __init__(
self, n_neurons, constraints, label, rate, max_rate, start,
duration, seed, max_atoms_per_core, model):
# pylint: disable=too-many-arguments
super(SpikeSourcePoissonVertex, self).__init__(
label, constraints, max_atoms_per_core)
# atoms params
self._n_atoms = n_neurons
self._model_name = "SpikeSourcePoisson"
self._model = model
self._seed = seed
self._kiss_seed = dict()
self._rng = None
self._n_subvertices = 0
self._n_data_specs = 0
# check for changes parameters
self._change_requires_mapping = True
self._change_requires_neuron_parameters_reload = False
# Prepare for recording, and to get spikes
self._spike_recorder = MultiSpikeRecorder()
# Store the parameters
self._max_rate = max_rate
self._rate = self.convert_rate(rate)
self._rate_change = numpy.zeros(self._rate.size)
self._start = utility_calls.convert_param_to_numpy(start, n_neurons)
self._duration = utility_calls.convert_param_to_numpy(
duration, n_neurons)
self._time_to_spike = utility_calls.convert_param_to_numpy(
0, n_neurons)
self._machine_time_step = None
# Prepare for recording, and to get spikes
self._spike_recorder = MultiSpikeRecorder()
@property
@overrides(AbstractChangableAfterRun.requires_mapping)
def requires_mapping(self):
return self._change_requires_mapping
@overrides(AbstractChangableAfterRun.mark_no_changes)
def mark_no_changes(self):
self._change_requires_mapping = False
@overrides(SimplePopulationSettable.set_value)
def set_value(self, key, value):
SimplePopulationSettable.set_value(self, key, value)
self._change_requires_neuron_parameters_reload = True
def _max_spikes_per_ts(self, machine_time_step):
ts_per_second = MICROSECONDS_PER_SECOND / float(machine_time_step)
if float(self._max_rate) / ts_per_second <= \
SLOW_RATE_PER_TICK_CUTOFF:
return 1
# experiement show at 1000 is result is typically higher than actual
chance_ts = 1000
max_spikes_per_ts = scipy.stats.poisson.ppf(
1.0 - (1.0 / float(chance_ts)),
float(self._max_rate) / ts_per_second)
return int(math.ceil(max_spikes_per_ts)) + 1.0
def get_recording_sdram_usage(self, vertex_slice, machine_time_step):
variable_sdram = self._spike_recorder.get_sdram_usage_in_bytes(
vertex_slice.n_atoms, self._max_spikes_per_ts(machine_time_step))
constant_sdram = ConstantSDRAM(
variable_sdram.per_timestep * OVERFLOW_TIMESTEPS_FOR_SDRAM)
return variable_sdram + constant_sdram
@inject_items({
"machine_time_step": "MachineTimeStep"
})
@overrides(
ApplicationVertex.get_resources_used_by_atoms,
additional_arguments={"machine_time_step"}
)
def get_resources_used_by_atoms(self, vertex_slice, machine_time_step):
# pylint: disable=arguments-differ
poisson_params_sz = self.get_params_bytes(vertex_slice)
other = ConstantSDRAM(
SYSTEM_BYTES_REQUIREMENT +
SpikeSourcePoissonMachineVertex.get_provenance_data_size(0) +
poisson_params_sz +
recording_utilities.get_recording_header_size(1) +
recording_utilities.get_recording_data_constant_size(1))
recording = self.get_recording_sdram_usage(
vertex_slice, machine_time_step)
# build resources as i currently know
container = ResourceContainer(
sdram=recording + other,
dtcm=DTCMResource(self.get_dtcm_usage_for_atoms()),
cpu_cycles=CPUCyclesPerTickResource(
self.get_cpu_usage_for_atoms()))
return container
@property
def n_atoms(self):
return self._n_atoms
def create_machine_vertex(
self, vertex_slice, resources_required, label=None,
constraints=None):
# pylint: disable=too-many-arguments, arguments-differ
self._n_subvertices += 1
return SpikeSourcePoissonMachineVertex(
resources_required, self._spike_recorder.record,
constraints, label)
@property
def rate(self):
return self._rate
def convert_rate(self, rate):
new_rates = utility_calls.convert_param_to_numpy(rate, self._n_atoms)
new_max = max(new_rates)
if self._max_rate is None:
self._max_rate = new_max
# Setting record forces reset so ok to go over if not recording
elif self._spike_recorder.record and new_max > self._max_rate:
logger.info('Increasing spike rate while recording requires a '
'"reset unless additional_parameters "max_rate" is '
'set')
self._change_requires_mapping = True
self._max_rate = new_max
return new_rates
@rate.setter
def rate(self, rate):
new_rate = self.convert_rate(rate)
self._rate_change = new_rate - self._rate
self._rate = new_rate
@property
def start(self):
return self._start
@start.setter
def start(self, start):
self._start = utility_calls.convert_param_to_numpy(
start, self._n_atoms)
@property
def duration(self):
return self._duration
@duration.setter
def duration(self, duration):
self._duration = utility_calls.convert_param_to_numpy(
duration, self._n_atoms)
@property
def seed(self):
return self._seed
@seed.setter
def seed(self, seed):
self._seed = seed
self._kiss_seed = dict()
self._rng = None
@staticmethod
def get_params_bytes(vertex_slice):
""" Gets the size of the poisson parameters in bytes
:param vertex_slice:
"""
return (PARAMS_BASE_WORDS +
(vertex_slice.n_atoms * PARAMS_WORDS_PER_NEURON)) * 4
def reserve_memory_regions(self, spec, placement, graph_mapper):
""" Reserve memory regions for poisson source parameters and output\
buffer.
:param spec: the data specification writer
:param placement: the location this vertex resides on in the machine
:param graph_mapper: the mapping between app and machine graphs
:return: None
"""
spec.comment("\nReserving memory space for data regions:\n\n")
# Reserve memory:
spec.reserve_memory_region(
region=_REGIONS.SYSTEM_REGION.value,
size=SIMULATION_N_BYTES,
label='setup')
# reserve poisson params dsg region
self._reserve_poisson_params_region(placement, graph_mapper, spec)
spec.reserve_memory_region(
region=_REGIONS.SPIKE_HISTORY_REGION.value,
size=recording_utilities.get_recording_header_size(1),
label="Recording")
placement.vertex.reserve_provenance_data_region(spec)
def _reserve_poisson_params_region(self, placement, graph_mapper, spec):
""" does the allocation for the poisson params region itself, as\
it can be reused for setters after an initial run
:param placement: the location on machine for this vertex
:param graph_mapper: the mapping between machine and application graphs
:param spec: the dsg writer
:return: None
"""
spec.reserve_memory_region(
region=_REGIONS.POISSON_PARAMS_REGION.value,
size=self.get_params_bytes(graph_mapper.get_slice(
placement.vertex)), label='PoissonParams')
def _write_poisson_parameters(
self, spec, graph, placement, routing_info,
vertex_slice, machine_time_step, time_scale_factor):
""" Generate Neuron Parameter data for Poisson spike sources
:param spec: the data specification writer
:param key: the routing key for this vertex
:param vertex_slice:\
the slice of atoms a machine vertex holds from its application\
vertex
:param machine_time_step: the time between timer tick updates.
:param time_scale_factor:\
the scaling between machine time step and real time
:return: None
"""
# pylint: disable=too-many-arguments, too-many-locals
spec.comment("\nWriting Neuron Parameters for {} poisson sources:\n"
.format(vertex_slice.n_atoms))
# Set the focus to the memory region 2 (neuron parameters):
spec.switch_write_focus(_REGIONS.POISSON_PARAMS_REGION.value)
# Write Key info for this core:
key = routing_info.get_first_key_from_pre_vertex(
placement.vertex, constants.SPIKE_PARTITION_ID)
spec.write_value(data=1 if key is not None else 0)
spec.write_value(data=key if key is not None else 0)
# Write the incoming mask if there is one
in_edges = graph.get_edges_ending_at_vertex_with_partition_name(
placement.vertex, constants.LIVE_POISSON_CONTROL_PARTITION_ID)
if len(in_edges) > 1:
raise ConfigurationException(
"Only one control edge can end at a Poisson vertex")
incoming_mask = 0
if len(in_edges) == 1:
in_edge = in_edges[0]
# Get the mask of the incoming keys
incoming_mask = \
routing_info.get_routing_info_for_edge(in_edge).first_mask
incoming_mask = ~incoming_mask & 0xFFFFFFFF
spec.write_value(incoming_mask)
# Write the offset value
max_offset = (
machine_time_step * time_scale_factor) // _MAX_OFFSET_DENOMINATOR
spec.write_value(
int(math.ceil(max_offset / self._n_subvertices)) *
self._n_data_specs)
self._n_data_specs += 1
# Write the number of microseconds between sending spikes
total_mean_rate = numpy.sum(self._rate)
if total_mean_rate > 0:
max_spikes = numpy.sum(scipy.stats.poisson.ppf(
1.0 - (1.0 / self._rate), self._rate))
spikes_per_timestep = (
max_spikes / (MICROSECONDS_PER_SECOND // machine_time_step))
# avoid a possible division by zero / small number (which may
# result in a value that doesn't fit in a uint32) by only
# setting time_between_spikes if spikes_per_timestep is > 1
time_between_spikes = 1.0
if spikes_per_timestep > 1:
time_between_spikes = (
(machine_time_step * time_scale_factor) /
(spikes_per_timestep * 2.0))
spec.write_value(data=int(time_between_spikes))
else:
# If the rate is 0 or less, set a "time between spikes" of 1
# to ensure that some time is put between spikes in event
# of a rate change later on
spec.write_value(data=1)
# Write the number of seconds per timestep (unsigned long fract)
spec.write_value(
data=float(machine_time_step) / MICROSECONDS_PER_SECOND,
data_type=DataType.U032)
# Write the number of timesteps per second (accum)
spec.write_value(
data=MICROSECONDS_PER_SECOND / float(machine_time_step),
data_type=DataType.S1615)
# Write the slow-rate-per-tick-cutoff (accum)
spec.write_value(
data=SLOW_RATE_PER_TICK_CUTOFF, data_type=DataType.S1615)
# Write the lo_atom id
spec.write_value(data=vertex_slice.lo_atom)
# Write the number of sources
spec.write_value(data=vertex_slice.n_atoms)
# Write the random seed (4 words), generated randomly!
kiss_key = (vertex_slice.lo_atom, vertex_slice.hi_atom)
if kiss_key not in self._kiss_seed:
if self._rng is None:
self._rng = numpy.random.RandomState(self._seed)
self._kiss_seed[kiss_key] = [
self._rng.randint(-0x80000000, 0x7FFFFFFF) + 0x80000000
for _ in range(4)]
for value in self._kiss_seed[kiss_key]:
spec.write_value(data=value)
# Compute the start times in machine time steps
start = self._start[vertex_slice.as_slice]
start_scaled = self._convert_ms_to_n_timesteps(
start, machine_time_step)
# Compute the end times as start times + duration in machine time steps
# (where duration is not None)
duration = self._duration[vertex_slice.as_slice]
end_scaled = numpy.zeros(len(duration), dtype="uint32")
none_positions = numpy.isnan(duration)
positions = numpy.invert(none_positions)
end_scaled[none_positions] = 0xFFFFFFFF
end_scaled[positions] = self._convert_ms_to_n_timesteps(
start[positions] + duration[positions], machine_time_step)
# Get the rates for the atoms
rates = self._rate[vertex_slice.as_slice].astype("float")
# Compute the spikes per tick for each atom
spikes_per_tick = (
rates * (float(machine_time_step) / MICROSECONDS_PER_SECOND))
# Determine which sources are fast and which are slow
is_fast_source = spikes_per_tick > SLOW_RATE_PER_TICK_CUTOFF
# Compute the e^-(spikes_per_tick) for fast sources to allow fast
# computation of the Poisson distribution to get the number of spikes
# per timestep
exp_minus_lambda = numpy.zeros(len(spikes_per_tick), dtype="float")
exp_minus_lambda[is_fast_source] = numpy.exp(
-1.0 * spikes_per_tick[is_fast_source])
# Compute the inter-spike-interval for slow sources to get the average
# number of timesteps between spikes
isi_val = numpy.zeros(len(spikes_per_tick), dtype="float")
elements = numpy.logical_not(is_fast_source) & (spikes_per_tick > 0)
isi_val[elements] = 1.0 / spikes_per_tick[elements]
# Get the time to spike value
time_to_spike = self._time_to_spike[vertex_slice.as_slice]
changed_rates = (
self._rate_change[vertex_slice.as_slice].astype("bool") & elements)
time_to_spike[changed_rates] = 0.0
# Merge the arrays as parameters per atom
data = numpy.dstack((
start_scaled.astype("uint32"),
end_scaled.astype("uint32"),
is_fast_source.astype("uint32"),
(exp_minus_lambda * (2 ** 32)).astype("uint32"),
(isi_val * (2 ** 15)).astype("uint32"),
(time_to_spike * (2 ** 15)).astype("uint32")
))[0]
spec.write_array(data)
@staticmethod
def _convert_ms_to_n_timesteps(value, machine_time_step):
return numpy.round(
value * (MICROSECONDS_PER_MILLISECOND / float(machine_time_step)))
@staticmethod
def _convert_n_timesteps_to_ms(value, machine_time_step):
return (
value / (MICROSECONDS_PER_MILLISECOND / float(machine_time_step)))
@overrides(AbstractSpikeRecordable.is_recording_spikes)
def is_recording_spikes(self):
return self._spike_recorder.record
@overrides(AbstractSpikeRecordable.set_recording_spikes)
def set_recording_spikes(
self, new_state=True, sampling_interval=None, indexes=None):
if sampling_interval is not None:
logger.warning("Sampling interval currently not supported for "
"SpikeSourcePoisson so being ignored")
if indexes is not None:
logger.warning("indexes not supported for "
"SpikeSourcePoisson so being ignored")
if new_state and not self._spike_recorder.record:
self._change_requires_mapping = True
self._spike_recorder.record = new_state
@overrides(AbstractSpikeRecordable.get_spikes_sampling_interval)
def get_spikes_sampling_interval(self):
return globals_variables.get_simulator().machine_time_step
@staticmethod
def get_dtcm_usage_for_atoms():
return 0
@staticmethod
def get_cpu_usage_for_atoms():
return 0
@inject_items({
"machine_time_step": "MachineTimeStep",
"time_scale_factor": "TimeScaleFactor",
"graph_mapper": "MemoryGraphMapper",
"routing_info": "MemoryRoutingInfos",
"graph": "MemoryMachineGraph"})
@overrides(
AbstractRewritesDataSpecification.regenerate_data_specification,
additional_arguments={
"machine_time_step", "time_scale_factor", "graph_mapper",
"routing_info", "graph"})
def regenerate_data_specification(
self, spec, placement, machine_time_step, time_scale_factor,
graph_mapper, routing_info, graph):
# pylint: disable=too-many-arguments, arguments-differ
# reserve the neuron parameters data region
self._reserve_poisson_params_region(placement, graph_mapper, spec)
# allocate parameters
self._write_poisson_parameters(
spec=spec, graph=graph, placement=placement,
routing_info=routing_info,
vertex_slice=graph_mapper.get_slice(placement.vertex),
machine_time_step=machine_time_step,
time_scale_factor=time_scale_factor)
# end spec
spec.end_specification()
@overrides(AbstractRewritesDataSpecification
.requires_memory_regions_to_be_reloaded)
def requires_memory_regions_to_be_reloaded(self):
return self._change_requires_neuron_parameters_reload
@overrides(AbstractRewritesDataSpecification.mark_regions_reloaded)
def mark_regions_reloaded(self):
self._change_requires_neuron_parameters_reload = False
@overrides(AbstractReadParametersBeforeSet.read_parameters_from_machine)
def read_parameters_from_machine(
self, transceiver, placement, vertex_slice):
# locate sdram address to where the neuron parameters are stored
poisson_parameter_region_sdram_address = \
helpful_functions.locate_memory_region_for_placement(
placement, _REGIONS.POISSON_PARAMS_REGION.value, transceiver)
# shift past the extra stuff before neuron parameters that we don't
# need to read
poisson_parameter_parameters_sdram_address = \
poisson_parameter_region_sdram_address + \
START_OF_POISSON_GENERATOR_PARAMETERS
# get size of poisson params
size_of_region = self.get_params_bytes(vertex_slice)
size_of_region -= START_OF_POISSON_GENERATOR_PARAMETERS
# get data from the machine
byte_array = transceiver.read_memory(
placement.x, placement.y,
poisson_parameter_parameters_sdram_address, size_of_region)
# Convert the data to parameter values
(start, end, is_fast_source, exp_minus_lambda, isi,
time_to_next_spike) = _PoissonStruct.read_data(
byte_array, 0, vertex_slice.n_atoms)
# Convert start values as timesteps into milliseconds
self._start[vertex_slice.as_slice] = self._convert_n_timesteps_to_ms(
start, self._machine_time_step)
# Convert end values as timesteps to durations in milliseconds
self._duration[vertex_slice.as_slice] = (
self._convert_n_timesteps_to_ms(end, self._machine_time_step) -
self._start[vertex_slice.as_slice])
# Work out the spikes per tick depending on if the source is slow
# or fast
is_fast_source = is_fast_source == 1.0
spikes_per_tick = numpy.zeros(len(is_fast_source), dtype="float")
spikes_per_tick[is_fast_source] = numpy.log(
exp_minus_lambda[is_fast_source]) * -1.0
slow_elements = isi > 0
spikes_per_tick[slow_elements] = 1.0 / isi[slow_elements]
# Convert spikes per tick to rates
self._rate[vertex_slice.as_slice] = (
spikes_per_tick *
(MICROSECONDS_PER_SECOND / float(self._machine_time_step)))
# Store the updated time until next spike so that it can be
# rewritten when the parameters are loaded
self._time_to_spike[vertex_slice.as_slice] = time_to_next_spike
@inject_items({
"machine_time_step": "MachineTimeStep",
"time_scale_factor": "TimeScaleFactor",
"graph_mapper": "MemoryGraphMapper",
"routing_info": "MemoryRoutingInfos",
"data_n_time_steps": "DataNTimeSteps",
"graph": "MemoryMachineGraph"
})
@overrides(
AbstractGeneratesDataSpecification.generate_data_specification,
additional_arguments={
"machine_time_step", "time_scale_factor", "graph_mapper",
"routing_info", "data_n_time_steps", "graph"
}
)
def generate_data_specification(
self, spec, placement, machine_time_step, time_scale_factor,
graph_mapper, routing_info, data_n_time_steps, graph):
# pylint: disable=too-many-arguments, arguments-differ
self._machine_time_step = machine_time_step
vertex = placement.vertex
vertex_slice = graph_mapper.get_slice(vertex)
spec.comment("\n*** Spec for SpikeSourcePoisson Instance ***\n\n")
# Reserve SDRAM space for memory areas:
self.reserve_memory_regions(spec, placement, graph_mapper)
# write setup data
spec.switch_write_focus(_REGIONS.SYSTEM_REGION.value)
spec.write_array(simulation_utilities.get_simulation_header_array(
self.get_binary_file_name(), machine_time_step,
time_scale_factor))
# write recording data
spec.switch_write_focus(_REGIONS.SPIKE_HISTORY_REGION.value)
sdram = self.get_recording_sdram_usage(
vertex_slice, machine_time_step)
recorded_region_sizes = [sdram.get_total_sdram(data_n_time_steps)]
spec.write_array(recording_utilities.get_recording_header_array(
recorded_region_sizes))
# write parameters
self._write_poisson_parameters(
spec, graph, placement, routing_info, vertex_slice,
machine_time_step, time_scale_factor)
# End-of-Spec:
spec.end_specification()
@overrides(AbstractHasAssociatedBinary.get_binary_file_name)
def get_binary_file_name(self):
return "spike_source_poisson.aplx"
@overrides(AbstractHasAssociatedBinary.get_binary_start_type)
def get_binary_start_type(self):
return ExecutableType.USES_SIMULATION_INTERFACE
@overrides(AbstractSpikeRecordable.get_spikes)
def get_spikes(
self, placements, graph_mapper, buffer_manager, machine_time_step):
return self._spike_recorder.get_spikes(
self.label, buffer_manager,
SpikeSourcePoissonVertex.SPIKE_RECORDING_REGION_ID,
placements, graph_mapper, self, machine_time_step)
@overrides(AbstractProvidesOutgoingPartitionConstraints.
get_outgoing_partition_constraints)
def get_outgoing_partition_constraints(self, partition):
return [ContiguousKeyRangeContraint()]
@overrides(AbstractSpikeRecordable.clear_spike_recording)
def clear_spike_recording(self, buffer_manager, placements, graph_mapper):
machine_vertices = graph_mapper.get_machine_vertices(self)
for machine_vertex in machine_vertices:
placement = placements.get_placement_of_vertex(machine_vertex)
buffer_manager.clear_recorded_data(
placement.x, placement.y, placement.p,
SpikeSourcePoissonVertex.SPIKE_RECORDING_REGION_ID)
def describe(self):
"""
Returns a human-readable description of the cell or synapse type.
The output may be customised by specifying a different template
together with an associated template engine
(see ``pyNN.descriptions``).
If template is None, then a dictionary containing the template context
will be returned.
"""
parameters = dict()
for parameter_name in self._model.default_parameters:
parameters[parameter_name] = self.get_value(parameter_name)
context = {
"name": self._model_name,
"default_parameters": self._model.default_parameters,
"default_initial_values": self._model.default_parameters,
"parameters": parameters,
}
return context
@property
def max_rate(self):
return self._max_rate
class SpikeSourcePoissonVertex(
ApplicationVertex, AbstractGeneratesDataSpecification,
AbstractHasAssociatedBinary, AbstractSpikeRecordable,
AbstractProvidesOutgoingPartitionConstraints,
AbstractChangableAfterRun, AbstractReadParametersBeforeSet,
AbstractRewritesDataSpecification, SimplePopulationSettable,
ProvidesKeyToAtomMappingImpl):
""" A Poisson Spike source object
"""
__slots__ = [
"__change_requires_mapping",
"__change_requires_neuron_parameters_reload", "__duration",
"__machine_time_step", "__model", "__model_name", "__n_atoms",
"__rate", "__rng", "__seed", "__spike_recorder", "__start",
"__time_to_spike", "__kiss_seed", "__n_subvertices", "__n_data_specs",
"__max_rate", "__rate_change", "__n_profile_samples", "__data",
"__is_variable_rate", "__max_spikes"
]
SPIKE_RECORDING_REGION_ID = 0
def __init__(self,
n_neurons,
constraints,
label,
seed,
max_atoms_per_core,
model,
rate=None,
start=None,
duration=None,
rates=None,
starts=None,
durations=None,
max_rate=None):
# pylint: disable=too-many-arguments
super(SpikeSourcePoissonVertex, self).__init__(label, constraints,
max_atoms_per_core)
# atoms params
self.__n_atoms = n_neurons
self.__model_name = "SpikeSourcePoisson"
self.__model = model
self.__seed = seed
self.__kiss_seed = dict()
self.__rng = None
self.__n_subvertices = 0
self.__n_data_specs = 0
# check for changes parameters
self.__change_requires_mapping = True
self.__change_requires_neuron_parameters_reload = False
self.__spike_recorder = MultiSpikeRecorder()
# Check for disallowed pairs of parameters
if (rates is not None) and (rate is not None):
raise Exception("Exactly one of rate and rates can be specified")
if (starts is not None) and (start is not None):
raise Exception("Exactly one of start and starts can be specified")
if (durations is not None) and (duration is not None):
raise Exception(
"Exactly one of duration and durations can be specified")
if rate is None and rates is None:
raise Exception("One of rate or rates must be specified")
# Normalise the parameters
self.__is_variable_rate = rates is not None
if rates is None:
if hasattr(rate, "__len__"):
# Single rate per neuron for whole simulation
rates = [numpy.array([r]) for r in rate]
else:
# Single rate for all neurons for whole simulation
rates = numpy.array([rate])
elif hasattr(rates[0], "__len__"):
# Convert each list to numpy array
rates = [numpy.array(r) for r in rates]
else:
rates = numpy.array(rates)
if starts is None and start is not None:
if hasattr(start, "__len__"):
starts = [numpy.array([s]) for s in start]
elif start is None:
starts = numpy.array([0])
else:
starts = numpy.array([start])
elif starts is not None and hasattr(starts[0], "__len__"):
starts = [numpy.array(s) for s in starts]
elif starts is not None:
starts = numpy.array(starts)
if durations is None and duration is not None:
if hasattr(duration, "__len__"):
durations = [numpy.array([d]) for d in duration]
else:
durations = numpy.array([duration])
elif durations is not None and hasattr(durations[0], "__len__"):
durations = [numpy.array(d) for d in durations]
elif durations is not None:
durations = numpy.array(durations)
else:
if hasattr(rates[0], "__len__"):
durations = [
numpy.array([None for r in _rate]) for _rate in rates
]
else:
durations = numpy.array([None for _rate in rates])
# Check that there is either one list for all neurons,
# or one per neuron
if hasattr(rates[0], "__len__") and len(rates) != n_neurons:
raise Exception(
"Must specify one rate for all neurons or one per neuron")
if (starts is not None and hasattr(starts[0], "__len__")
and len(starts) != n_neurons):
raise Exception(
"Must specify one start for all neurons or one per neuron")
if (durations is not None and hasattr(durations[0], "__len__")
and len(durations) != n_neurons):
raise Exception(
"Must specify one duration for all neurons or one per neuron")
# Check that for each rate there is a start and duration if needed
# TODO: Could be more efficient for case where parameters are not one
# per neuron
for i in range(n_neurons):
rate_set = rates
if hasattr(rates[0], "__len__"):
rate_set = rates[i]
if not hasattr(rate_set, "__len__"):
raise Exception("Multiple rates must be a list")
if starts is None and len(rate_set) > 1:
raise Exception("When multiple rates are specified,"
" each must have a start")
elif starts is not None:
start_set = starts
if hasattr(starts[0], "__len__"):
start_set = starts[i]
if len(start_set) != len(rate_set):
raise Exception("Each rate must have a start")
if any(s is None for s in start_set):
raise Exception("Start must not be None")
if durations is not None:
duration_set = durations
if hasattr(durations[0], "__len__"):
duration_set = durations[i]
if len(duration_set) != len(rate_set):
raise Exception("Each rate must have its own duration")
if hasattr(rates[0], "__len__"):
time_to_spike = [
numpy.array([0 for _ in range(len(rates[i]))])
for i in range(len(rates))
]
else:
time_to_spike = numpy.array([0 for _ in range(len(rates))])
self.__data = SpynnakerRangeDictionary(n_neurons)
self.__data["rates"] = SpynnakerRangedList(
n_neurons,
rates,
use_list_as_value=not hasattr(rates[0], "__len__"))
self.__data["starts"] = SpynnakerRangedList(
n_neurons,
starts,
use_list_as_value=not hasattr(starts[0], "__len__"))
self.__data["durations"] = SpynnakerRangedList(
n_neurons,
durations,
use_list_as_value=not hasattr(durations[0], "__len__"))
self.__data["time_to_spike"] = SpynnakerRangedList(
n_neurons,
time_to_spike,
use_list_as_value=not hasattr(time_to_spike[0], "__len__"))
self.__rng = numpy.random.RandomState(seed)
self.__rate_change = numpy.zeros(n_neurons)
self.__machine_time_step = None
# get config from simulator
config = globals_variables.get_simulator().config
self.__n_profile_samples = helpful_functions.read_config_int(
config, "Reports", "n_profile_samples")
# Prepare for recording, and to get spikes
self.__spike_recorder = MultiSpikeRecorder()
all_rates = list(_flatten(self.__data["rates"]))
self.__max_rate = max_rate
if max_rate is None and len(all_rates):
self.__max_rate = numpy.amax(all_rates)
elif max_rate is None:
self.__max_rate = 0
total_rate = numpy.sum(all_rates)
self.__max_spikes = 0
if total_rate > 0:
# Note we have to do this per rate, as the whole array is not numpy
max_rates = numpy.array(
[numpy.max(r) for r in self.__data["rates"]])
self.__max_spikes = numpy.sum(
scipy.stats.poisson.ppf(1.0 - (1.0 / max_rates), max_rates))
@property
def rate(self):
if self.__is_variable_rate:
raise Exception("Get variable rate poisson rates with .rates")
return list(_flatten(self.__data["rates"]))
@rate.setter
def rate(self, rate):
if self.__is_variable_rate:
raise Exception("Cannot set rate of a variable rate poisson")
self.__rate_change = rate - numpy.array(
list(_flatten(self.__data["rates"])))
# Normalise parameter
if hasattr(rate, "__len__"):
# Single rate per neuron for whole simulation
self.__data["rates"].set_value([numpy.array([r]) for r in rate])
else:
# Single rate for all neurons for whole simulation
self.__data["rates"].set_value(numpy.array([rate]),
use_list_as_value=True)
all_rates = list(_flatten(self.__data["rates"]))
new_max = 0
if len(all_rates):
new_max = numpy.amax(all_rates)
if self.__max_rate is None:
self.__max_rate = new_max
# Setting record forces reset so OK to go over if not recording
elif self.__spike_recorder.record and new_max > self.__max_rate:
logger.info('Increasing spike rate while recording requires a '
'"reset unless additional_parameters "max_rate" is '
'set')
self.__change_requires_mapping = True
self.__max_rate = new_max
@property
def start(self):
if self.__is_variable_rate:
raise Exception("Get variable rate poisson starts with .starts")
return self.__data["starts"]
@start.setter
def start(self, start):
if self.__is_variable_rate:
raise Exception("Cannot set start of a variable rate poisson")
# Normalise parameter
if hasattr(start, "__len__"):
# Single start per neuron for whole simulation
self.__data["starts"].set_value([numpy.array([s]) for s in start])
else:
# Single start for all neurons for whole simulation
self.__data["starts"].set_value(numpy.array([start]),
use_list_as_value=True)
@property
def duration(self):
if self.__is_variable_rate:
raise Exception(
"Get variable rate poisson durations with .durations")
return self.__data["durations"]
@duration.setter
def duration(self, duration):
if self.__is_variable_rate:
raise Exception("Cannot set duration of a variable rate poisson")
# Normalise parameter
if hasattr(duration, "__len__"):
# Single duration per neuron for whole simulation
self.__data["durations"].set_value(
[numpy.array([d]) for d in duration])
else:
# Single duration for all neurons for whole simulation
self.__data["durations"].set_value(numpy.array([duration]),
use_list_as_value=True)
@property
def rates(self):
return self.__data["rates"]
@rates.setter
def rates(self, _rates):
if self.__is_variable_rate:
raise Exception("Cannot set rates of a variable rate poisson")
raise Exception("Set the rate of a Poisson source using rate")
@property
def starts(self):
return self.__data["starts"]
@starts.setter
def starts(self, _starts):
if self.__is_variable_rate:
raise Exception("Cannot set starts of a variable rate poisson")
raise Exception("Set the start of a Poisson source using start")
@property
def durations(self):
return self.__data["durations"]
@durations.setter
def durations(self, _durations):
if self.__is_variable_rate:
raise Exception("Cannot set durations of a variable rate poisson")
raise Exception("Set the duration of a Poisson source using duration")
@property
@overrides(AbstractChangableAfterRun.requires_mapping)
def requires_mapping(self):
return self.__change_requires_mapping
@overrides(AbstractChangableAfterRun.mark_no_changes)
def mark_no_changes(self):
self.__change_requires_mapping = False
@overrides(SimplePopulationSettable.set_value)
def set_value(self, key, value):
SimplePopulationSettable.set_value(self, key, value)
self.__change_requires_neuron_parameters_reload = True
def _max_spikes_per_ts(self, machine_time_step):
ts_per_second = MICROSECONDS_PER_SECOND / float(machine_time_step)
if float(self.__max_rate) / ts_per_second < \
SLOW_RATE_PER_TICK_CUTOFF:
return 1
# Experiments show at 1000 this result is typically higher than actual
chance_ts = 1000
max_spikes_per_ts = scipy.stats.poisson.ppf(
1.0 - (1.0 / float(chance_ts)),
float(self.__max_rate) / ts_per_second)
return int(math.ceil(max_spikes_per_ts)) + 1.0
def get_recording_sdram_usage(self, vertex_slice, machine_time_step):
variable_sdram = self.__spike_recorder.get_sdram_usage_in_bytes(
vertex_slice.n_atoms, self._max_spikes_per_ts(machine_time_step))
constant_sdram = ConstantSDRAM(variable_sdram.per_timestep *
OVERFLOW_TIMESTEPS_FOR_SDRAM)
return variable_sdram + constant_sdram
@inject_items({"machine_time_step": "MachineTimeStep"})
@overrides(ApplicationVertex.get_resources_used_by_atoms,
additional_arguments={"machine_time_step"})
def get_resources_used_by_atoms(self, vertex_slice, machine_time_step):
# pylint: disable=arguments-differ
poisson_params_sz = self.get_rates_bytes(vertex_slice)
other = ConstantSDRAM(
SYSTEM_BYTES_REQUIREMENT +
SpikeSourcePoissonMachineVertex.get_provenance_data_size(0) +
poisson_params_sz +
recording_utilities.get_recording_header_size(1) +
recording_utilities.get_recording_data_constant_size(1) +
profile_utils.get_profile_region_size(self.__n_profile_samples))
recording = self.get_recording_sdram_usage(vertex_slice,
machine_time_step)
# build resources as i currently know
container = ResourceContainer(sdram=recording + other,
dtcm=DTCMResource(
self.get_dtcm_usage_for_atoms()),
cpu_cycles=CPUCyclesPerTickResource(
self.get_cpu_usage_for_atoms()))
return container
@property
def n_atoms(self):
return self.__n_atoms
def create_machine_vertex(self,
vertex_slice,
resources_required,
label=None,
constraints=None):
# pylint: disable=too-many-arguments, arguments-differ
self.__n_subvertices += 1
return SpikeSourcePoissonMachineVertex(resources_required,
self.__spike_recorder.record,
constraints, label)
@property
def max_rate(self):
return self.__max_rate
@property
def seed(self):
return self.__seed
@seed.setter
def seed(self, seed):
self.__seed = seed
self.__kiss_seed = dict()
self.__rng = None
def get_rates_bytes(self, vertex_slice):
""" Gets the size of the Poisson rates in bytes
:param vertex_slice:
"""
n_rates = sum(
len(self.__data["rates"][i])
for i in range(vertex_slice.lo_atom, vertex_slice.hi_atom + 1))
return ((vertex_slice.n_atoms * PARAMS_WORDS_PER_NEURON) +
(n_rates * PARAMS_WORDS_PER_RATE)) * BYTES_PER_WORD
def reserve_memory_regions(self, spec, placement, graph_mapper):
""" Reserve memory regions for Poisson source parameters and output\
buffer.
:param spec: the data specification writer
:param placement: the location this vertex resides on in the machine
:param graph_mapper: the mapping between app and machine graphs
:return: None
"""
spec.comment("\nReserving memory space for data regions:\n\n")
# Reserve memory:
spec.reserve_memory_region(region=_REGIONS.SYSTEM_REGION.value,
size=SIMULATION_N_BYTES,
label='setup')
# reserve poisson parameters and rates DSG region
self._reserve_poisson_params_rates_region(placement, graph_mapper,
spec)
spec.reserve_memory_region(
region=_REGIONS.SPIKE_HISTORY_REGION.value,
size=recording_utilities.get_recording_header_size(1),
label="Recording")
profile_utils.reserve_profile_region(spec,
_REGIONS.PROFILER_REGION.value,
self.__n_profile_samples)
placement.vertex.reserve_provenance_data_region(spec)
def _reserve_poisson_params_rates_region(self, placement, graph_mapper,
spec):
""" Allocate space for the Poisson parameters and rates regions as\
they can be reused for setters after an initial run
:param placement: the location on machine for this vertex
:param graph_mapper: the mapping between machine and application graphs
:param spec: the DSG writer
:return: None
"""
spec.reserve_memory_region(region=_REGIONS.POISSON_PARAMS_REGION.value,
size=PARAMS_BASE_WORDS * 4,
label="PoissonParams")
spec.reserve_memory_region(region=_REGIONS.RATES_REGION.value,
size=self.get_rates_bytes(
graph_mapper.get_slice(
placement.vertex)),
label='PoissonRates')
def _write_poisson_parameters(self, spec, graph, placement, routing_info,
vertex_slice, machine_time_step,
time_scale_factor):
""" Generate Parameter data for Poisson spike sources
:param spec: the data specification writer
:param key: the routing key for this vertex
:param vertex_slice:\
the slice of atoms a machine vertex holds from its application\
vertex
:param machine_time_step: the time between timer tick updates.
:param time_scale_factor:\
the scaling between machine time step and real time
"""
# pylint: disable=too-many-arguments, too-many-locals
spec.comment("\nWriting Parameters for {} poisson sources:\n".format(
vertex_slice.n_atoms))
# Set the focus to the memory region 2 (neuron parameters):
spec.switch_write_focus(_REGIONS.POISSON_PARAMS_REGION.value)
# Write Key info for this core:
key = routing_info.get_first_key_from_pre_vertex(
placement.vertex, constants.SPIKE_PARTITION_ID)
spec.write_value(data=1 if key is not None else 0)
spec.write_value(data=key if key is not None else 0)
# Write the incoming mask if there is one
in_edges = graph.get_edges_ending_at_vertex_with_partition_name(
placement.vertex, constants.LIVE_POISSON_CONTROL_PARTITION_ID)
if len(in_edges) > 1:
raise ConfigurationException(
"Only one control edge can end at a Poisson vertex")
incoming_mask = 0
if len(in_edges) == 1:
in_edge = in_edges[0]
# Get the mask of the incoming keys
incoming_mask = \
routing_info.get_routing_info_for_edge(in_edge).first_mask
incoming_mask = ~incoming_mask & 0xFFFFFFFF
spec.write_value(incoming_mask)
# Write the offset value
max_offset = (machine_time_step *
time_scale_factor) // _MAX_OFFSET_DENOMINATOR
spec.write_value(
int(math.ceil(max_offset / self.__n_subvertices)) *
self.__n_data_specs)
self.__n_data_specs += 1
if self.__max_spikes > 0:
spikes_per_timestep = (
self.__max_spikes /
(MICROSECONDS_PER_SECOND // machine_time_step))
# avoid a possible division by zero / small number (which may
# result in a value that doesn't fit in a uint32) by only
# setting time_between_spikes if spikes_per_timestep is > 1
time_between_spikes = 1.0
if spikes_per_timestep > 1:
time_between_spikes = (
(machine_time_step * time_scale_factor) /
(spikes_per_timestep * 2.0))
spec.write_value(data=int(time_between_spikes))
else:
# If the rate is 0 or less, set a "time between spikes" of 1
# to ensure that some time is put between spikes in event
# of a rate change later on
spec.write_value(data=1)
# Write the number of seconds per timestep (unsigned long fract)
spec.write_value(data=float(machine_time_step) /
MICROSECONDS_PER_SECOND,
data_type=DataType.U032)
# Write the number of timesteps per second (integer)
spec.write_value(data=int(MICROSECONDS_PER_SECOND /
float(machine_time_step)))
# Write the slow-rate-per-tick-cutoff (accum)
spec.write_value(data=SLOW_RATE_PER_TICK_CUTOFF,
data_type=DataType.S1615)
# Write the fast-rate-per-tick-cutoff (accum)
spec.write_value(data=FAST_RATE_PER_TICK_CUTOFF,
data_type=DataType.S1615)
# Write the lo_atom ID
spec.write_value(data=vertex_slice.lo_atom)
# Write the number of sources
spec.write_value(data=vertex_slice.n_atoms)
# Write the random seed (4 words), generated randomly!
kiss_key = (vertex_slice.lo_atom, vertex_slice.hi_atom)
if kiss_key not in self.__kiss_seed:
if self.__rng is None:
self.__rng = numpy.random.RandomState(self.__seed)
self.__kiss_seed[kiss_key] = validate_mars_kiss_64_seed([
self.__rng.randint(-0x80000000, 0x7FFFFFFF) + 0x80000000
for _ in range(4)
])
for value in self.__kiss_seed[kiss_key]:
spec.write_value(data=value)
def _write_poisson_rates(self, spec, vertex_slice, machine_time_step,
first_machine_time_step):
""" Generate Rate data for Poisson spike sources
:param spec: the data specification writer
:param vertex_slice:\
the slice of atoms a machine vertex holds from its application\
vertex
:param machine_time_step: the time between timer tick updates.
:param first_machine_time_step:\
First machine time step to start from the correct index
"""
spec.comment("\nWriting Rates for {} poisson sources:\n".format(
vertex_slice.n_atoms))
# Set the focus to the memory region 2 (neuron parameters):
spec.switch_write_focus(_REGIONS.RATES_REGION.value)
# Extract the data on which to work and convert to appropriate form
starts = numpy.array(
list(_flatten(
self.__data["starts"][vertex_slice.as_slice]))).astype("float")
durations = numpy.array(
list(_flatten(self.__data["durations"][
vertex_slice.as_slice]))).astype("float")
local_rates = self.__data["rates"][vertex_slice.as_slice]
n_rates = numpy.array([len(r) for r in local_rates])
splits = numpy.cumsum(n_rates)
rates = numpy.array(list(_flatten(local_rates)))
time_to_spike = numpy.array(
list(_flatten(self.__data["time_to_spike"][
vertex_slice.as_slice]))).astype("u4")
rate_change = self.__rate_change[vertex_slice.as_slice]
# Convert start times to start time steps
starts_scaled = self._convert_ms_to_n_timesteps(
starts, machine_time_step)
# Convert durations to end time steps, using the maximum for "None"
# duration (which means "until the end")
no_duration = numpy.isnan(durations)
durations_filtered = numpy.where(no_duration, 0, durations)
ends_scaled = self._convert_ms_to_n_timesteps(
durations_filtered, machine_time_step) + starts_scaled
ends_scaled = numpy.where(no_duration, _MAX_TIMESTEP, ends_scaled)
# Work out the timestep at which the next rate activates, using
# the maximum value at the end (meaning there is no "next")
starts_split = numpy.array_split(starts_scaled, splits)
next_scaled = numpy.concatenate(
[numpy.append(s[1:], _MAX_TIMESTEP) for s in starts_split[:-1]])
# Compute the spikes per tick for each rate for each atom
spikes_per_tick = rates * (float(machine_time_step) /
MICROSECONDS_PER_SECOND)
# Determine the properties of the sources
is_fast_source = spikes_per_tick >= SLOW_RATE_PER_TICK_CUTOFF
is_faster_source = spikes_per_tick >= FAST_RATE_PER_TICK_CUTOFF
not_zero = spikes_per_tick > 0
# pylint: disable=assignment-from-no-return
is_slow_source = numpy.logical_not(is_fast_source)
# Compute the e^-(spikes_per_tick) for fast sources to allow fast
# computation of the Poisson distribution to get the number of
# spikes per timestep
exp_minus_lambda = DataType.U032.encode_as_numpy_int_array(
numpy.where(is_fast_source, numpy.exp(-1.0 * spikes_per_tick), 0))
# Compute sqrt(lambda) for "faster" sources to allow Gaussian
# approximation of the Poisson distribution to get the number of
# spikes per timestep
sqrt_lambda = DataType.S1615.encode_as_numpy_int_array(
numpy.where(is_faster_source, numpy.sqrt(spikes_per_tick), 0))
# Compute the inter-spike-interval for slow sources to get the
# average number of timesteps between spikes
isi_val = numpy.where(not_zero & is_slow_source,
(1.0 / spikes_per_tick).astype(int),
0).astype("uint32")
# Reuse the time-to-spike read from the machine (if has been run)
# or don't if the rate has since been changed
time_to_spike_split = numpy.array_split(time_to_spike, splits)
time_to_spike = numpy.concatenate([
t if rate_change[i] else numpy.repeat(0, len(t))
for i, t in enumerate(time_to_spike_split[:-1])
])
# Turn the fast source booleans into uint32
is_fast_source = is_fast_source.astype("uint32")
# Group together the rate data for the core by rate
core_data = numpy.dstack(
(starts_scaled, ends_scaled, next_scaled, is_fast_source,
exp_minus_lambda, sqrt_lambda, isi_val, time_to_spike))[0]
# Group data by neuron id
core_data_split = numpy.array_split(core_data, splits)
# Work out the index where the core should start based on the given
# first timestep
ends_scaled_split = numpy.array_split(ends_scaled, splits)
indices = [
numpy.argmax(e > first_machine_time_step)
for e in ends_scaled_split[:-1]
]
# Build the final data for this core, and write it
final_data = numpy.concatenate([
numpy.concatenate(([len(d), indices[i]], numpy.concatenate(d)))
for i, d in enumerate(core_data_split[:-1])
])
spec.write_array(final_data)
@staticmethod
def _convert_ms_to_n_timesteps(value, machine_time_step):
return numpy.round(value * (MICROSECONDS_PER_MILLISECOND /
float(machine_time_step))).astype("uint32")
@staticmethod
def _convert_n_timesteps_to_ms(value, machine_time_step):
return (value /
(MICROSECONDS_PER_MILLISECOND / float(machine_time_step)))
@overrides(AbstractSpikeRecordable.is_recording_spikes)
def is_recording_spikes(self):
return self.__spike_recorder.record
@overrides(AbstractSpikeRecordable.set_recording_spikes)
def set_recording_spikes(self,
new_state=True,
sampling_interval=None,
indexes=None):
if sampling_interval is not None:
logger.warning("Sampling interval currently not supported for "
"SpikeSourcePoisson so being ignored")
if indexes is not None:
logger.warning("indexes not supported for "
"SpikeSourcePoisson so being ignored")
if new_state and not self.__spike_recorder.record:
self.__change_requires_mapping = True
self.__spike_recorder.record = new_state
@overrides(AbstractSpikeRecordable.get_spikes_sampling_interval)
def get_spikes_sampling_interval(self):
return globals_variables.get_simulator().machine_time_step
@staticmethod
def get_dtcm_usage_for_atoms():
return 0
@staticmethod
def get_cpu_usage_for_atoms():
return 0
@inject_items({
"machine_time_step": "MachineTimeStep",
"time_scale_factor": "TimeScaleFactor",
"graph_mapper": "MemoryGraphMapper",
"routing_info": "MemoryRoutingInfos",
"graph": "MemoryMachineGraph",
"first_machine_time_step": "FirstMachineTimeStep"
})
@overrides(AbstractRewritesDataSpecification.regenerate_data_specification,
additional_arguments={
"machine_time_step", "time_scale_factor", "graph_mapper",
"routing_info", "graph", "first_machine_time_step"
})
def regenerate_data_specification(self, spec, placement, machine_time_step,
time_scale_factor, graph_mapper,
routing_info, graph,
first_machine_time_step):
# pylint: disable=too-many-arguments, arguments-differ
# reserve the neuron parameters data region
self._reserve_poisson_params_rates_region(placement, graph_mapper,
spec)
# write parameters
vertex_slice = graph_mapper.get_slice(placement.vertex)
self._write_poisson_parameters(spec=spec,
graph=graph,
placement=placement,
routing_info=routing_info,
vertex_slice=vertex_slice,
machine_time_step=machine_time_step,
time_scale_factor=time_scale_factor)
# write rates
self._write_poisson_rates(spec, vertex_slice, machine_time_step,
first_machine_time_step)
# end spec
spec.end_specification()
@inject_items({"first_machine_time_step": "FirstMachineTimeStep"})
@overrides(AbstractRewritesDataSpecification.
requires_memory_regions_to_be_reloaded,
additional_arguments={"first_machine_time_step"})
def requires_memory_regions_to_be_reloaded(self, first_machine_time_step):
# pylint: disable=arguments-differ
return (self.__change_requires_neuron_parameters_reload
or first_machine_time_step == 0)
@overrides(AbstractRewritesDataSpecification.mark_regions_reloaded)
def mark_regions_reloaded(self):
self.__change_requires_neuron_parameters_reload = False
@overrides(AbstractReadParametersBeforeSet.read_parameters_from_machine)
def read_parameters_from_machine(self, transceiver, placement,
vertex_slice):
# locate SDRAM address where parameters are stored
poisson_params = \
helpful_functions.locate_memory_region_for_placement(
placement, _REGIONS.POISSON_PARAMS_REGION.value, transceiver)
seed_array = transceiver.read_memory(
placement.x, placement.y, poisson_params + SEED_OFFSET_BYTES,
SEED_SIZE_BYTES)
kiss_key = (vertex_slice.lo_atom, vertex_slice.hi_atom)
self.__kiss_seed[kiss_key] = struct.unpack_from("<4I", seed_array)
# locate SDRAM address where the rates are stored
poisson_rate_region_sdram_address = \
helpful_functions.locate_memory_region_for_placement(
placement, _REGIONS.RATES_REGION.value, transceiver)
# get size of poisson params
size_of_region = self.get_rates_bytes(vertex_slice)
# get data from the machine
byte_array = transceiver.read_memory(
placement.x, placement.y, poisson_rate_region_sdram_address,
size_of_region)
# For each atom, read the number of rates and the rate parameters
offset = 0
for i in range(vertex_slice.lo_atom, vertex_slice.hi_atom + 1):
n_values = struct.unpack_from("<I", byte_array, offset)[0]
offset += 4
# Skip reading the index, as it will be recalculated on data write
offset += 4
(_start, _end, _next, is_fast_source, exp_minus_lambda,
sqrt_lambda, isi, time_to_next_spike) = _PoissonStruct.read_data(
byte_array, offset, n_values)
offset += _PoissonStruct.get_size_in_whole_words(n_values) * 4
# Work out the spikes per tick depending on if the source is
# slow (isi), fast (exp) or faster (sqrt)
is_fast_source = is_fast_source == 1.0
spikes_per_tick = numpy.zeros(len(is_fast_source), dtype="float")
spikes_per_tick[is_fast_source] = numpy.log(
exp_minus_lambda[is_fast_source]) * -1.0
is_faster_source = sqrt_lambda > 0
# pylint: disable=assignment-from-no-return
spikes_per_tick[is_faster_source] = numpy.square(
sqrt_lambda[is_faster_source])
slow_elements = isi > 0
spikes_per_tick[slow_elements] = 1.0 / isi[slow_elements]
# Convert spikes per tick to rates
self.__data["rates"].set_value_by_id(
i,
spikes_per_tick *
(MICROSECONDS_PER_SECOND / float(self.__machine_time_step)))
# Store the updated time until next spike so that it can be
# rewritten when the parameters are loaded
self.__data["time_to_spike"].set_value_by_id(i, time_to_next_spike)
@inject_items({
"machine_time_step": "MachineTimeStep",
"time_scale_factor": "TimeScaleFactor",
"graph_mapper": "MemoryGraphMapper",
"routing_info": "MemoryRoutingInfos",
"data_n_time_steps": "DataNTimeSteps",
"graph": "MemoryMachineGraph",
"first_machine_time_step": "FirstMachineTimeStep"
})
@overrides(AbstractGeneratesDataSpecification.generate_data_specification,
additional_arguments={
"machine_time_step", "time_scale_factor", "graph_mapper",
"routing_info", "data_n_time_steps", "graph",
"first_machine_time_step"
})
def generate_data_specification(self, spec, placement, machine_time_step,
time_scale_factor, graph_mapper,
routing_info, data_n_time_steps, graph,
first_machine_time_step):
# pylint: disable=too-many-arguments, arguments-differ
self.__machine_time_step = machine_time_step
vertex = placement.vertex
vertex_slice = graph_mapper.get_slice(vertex)
spec.comment("\n*** Spec for SpikeSourcePoisson Instance ***\n\n")
# Reserve SDRAM space for memory areas:
self.reserve_memory_regions(spec, placement, graph_mapper)
# write setup data
spec.switch_write_focus(_REGIONS.SYSTEM_REGION.value)
spec.write_array(
simulation_utilities.get_simulation_header_array(
self.get_binary_file_name(), machine_time_step,
time_scale_factor))
# write recording data
spec.switch_write_focus(_REGIONS.SPIKE_HISTORY_REGION.value)
sdram = self.get_recording_sdram_usage(vertex_slice, machine_time_step)
recorded_region_sizes = [sdram.get_total_sdram(data_n_time_steps)]
spec.write_array(
recording_utilities.get_recording_header_array(
recorded_region_sizes))
# write parameters
self._write_poisson_parameters(spec, graph, placement, routing_info,
vertex_slice, machine_time_step,
time_scale_factor)
# write rates
self._write_poisson_rates(spec, vertex_slice, machine_time_step,
first_machine_time_step)
# write profile data
profile_utils.write_profile_region_data(spec,
_REGIONS.PROFILER_REGION.value,
self.__n_profile_samples)
# End-of-Spec:
spec.end_specification()
@overrides(AbstractHasAssociatedBinary.get_binary_file_name)
def get_binary_file_name(self):
return "spike_source_poisson.aplx"
@overrides(AbstractHasAssociatedBinary.get_binary_start_type)
def get_binary_start_type(self):
return ExecutableType.USES_SIMULATION_INTERFACE
@overrides(AbstractSpikeRecordable.get_spikes)
def get_spikes(self, placements, graph_mapper, buffer_manager,
machine_time_step):
return self.__spike_recorder.get_spikes(
self.label, buffer_manager,
SpikeSourcePoissonVertex.SPIKE_RECORDING_REGION_ID, placements,
graph_mapper, self, machine_time_step)
@overrides(AbstractProvidesOutgoingPartitionConstraints.
get_outgoing_partition_constraints)
def get_outgoing_partition_constraints(self, partition):
return [ContiguousKeyRangeContraint()]
@overrides(AbstractSpikeRecordable.clear_spike_recording)
def clear_spike_recording(self, buffer_manager, placements, graph_mapper):
machine_vertices = graph_mapper.get_machine_vertices(self)
for machine_vertex in machine_vertices:
placement = placements.get_placement_of_vertex(machine_vertex)
buffer_manager.clear_recorded_data(
placement.x, placement.y, placement.p,
SpikeSourcePoissonVertex.SPIKE_RECORDING_REGION_ID)
def describe(self):
""" Return a human-readable description of the cell or synapse type.
The output may be customised by specifying a different template\
together with an associated template engine\
(see ``pyNN.descriptions``).
If template is None, then a dictionary containing the template context\
will be returned.
"""
parameters = dict()
for parameter_name in self.__model.default_parameters:
parameters[parameter_name] = self.get_value(parameter_name)
context = {
"name": self.__model_name,
"default_parameters": self.__model.default_parameters,
"default_initial_values": self.__model.default_parameters,
"parameters": parameters,
}
return context
class SpikeSourcePoissonVertex(
ApplicationVertex, AbstractGeneratesDataSpecification,
AbstractHasAssociatedBinary, AbstractSpikeRecordable,
AbstractProvidesOutgoingPartitionConstraints,
AbstractChangableAfterRun, AbstractReadParametersBeforeSet,
AbstractRewritesDataSpecification, SimplePopulationSettable,
ProvidesKeyToAtomMappingImpl):
""" A Poisson Spike source object
"""
_N_POPULATION_RECORDING_REGIONS = 1
_DEFAULT_MALLOCS_USED = 2
SPIKE_RECORDING_REGION_ID = 0
# A count of the number of poisson vertices, to work out the random
# back off range
_n_poisson_machine_vertices = 0
def __init__(self, n_neurons, constraints, label, rate, start, duration,
seed, max_atoms_per_core, model):
# pylint: disable=too-many-arguments
super(SpikeSourcePoissonVertex, self).__init__(label, constraints,
max_atoms_per_core)
config = globals_variables.get_simulator().config
# atoms params
self._n_atoms = n_neurons
self._model_name = "SpikeSourcePoisson"
self._model = model
self._seed = seed
self._kiss_seed = dict()
self._rng = None
# check for changes parameters
self._change_requires_mapping = True
self._change_requires_neuron_parameters_reload = False
# Store the parameters
self._rate = utility_calls.convert_param_to_numpy(rate, n_neurons)
self._rate_change = numpy.zeros(self._rate.size)
self._start = utility_calls.convert_param_to_numpy(start, n_neurons)
self._duration = utility_calls.convert_param_to_numpy(
duration, n_neurons)
self._time_to_spike = utility_calls.convert_param_to_numpy(
0, n_neurons)
self._machine_time_step = None
# Prepare for recording, and to get spikes
self._spike_recorder = MultiSpikeRecorder()
self._time_between_requests = config.getint("Buffers",
"time_between_requests")
self._receive_buffer_host = config.get("Buffers",
"receive_buffer_host")
self._receive_buffer_port = helpful_functions.read_config_int(
config, "Buffers", "receive_buffer_port")
self._minimum_buffer_sdram = config.getint("Buffers",
"minimum_buffer_sdram")
self._using_auto_pause_and_resume = config.getboolean(
"Buffers", "use_auto_pause_and_resume")
spike_buffer_max_size = 0
self._buffer_size_before_receive = None
if config.getboolean("Buffers", "enable_buffered_recording"):
spike_buffer_max_size = config.getint("Buffers",
"spike_buffer_size")
self._buffer_size_before_receive = config.getint(
"Buffers", "buffer_size_before_receive")
self._maximum_sdram_for_buffering = [spike_buffer_max_size]
@property
@overrides(AbstractChangableAfterRun.requires_mapping)
def requires_mapping(self):
return self._change_requires_mapping
@overrides(AbstractChangableAfterRun.mark_no_changes)
def mark_no_changes(self):
self._change_requires_mapping = False
@overrides(SimplePopulationSettable.set_value)
def set_value(self, key, value):
SimplePopulationSettable.set_value(self, key, value)
self._change_requires_neuron_parameters_reload = True
def _max_spikes_per_ts(self, vertex_slice, n_machine_time_steps,
machine_time_step):
max_rate = numpy.amax(self._rate[vertex_slice.as_slice])
if max_rate == 0:
return 0
ts_per_second = MICROSECONDS_PER_SECOND / float(machine_time_step)
max_spikes_per_ts = scipy.stats.poisson.ppf(
1.0 - (1.0 / float(n_machine_time_steps)),
float(max_rate) / ts_per_second)
return int(math.ceil(max_spikes_per_ts)) + 1.0
@inject_items({
"n_machine_time_steps": "TotalMachineTimeSteps",
"machine_time_step": "MachineTimeStep"
})
@overrides(
ApplicationVertex.get_resources_used_by_atoms,
additional_arguments={"n_machine_time_steps", "machine_time_step"})
def get_resources_used_by_atoms(self, vertex_slice, n_machine_time_steps,
machine_time_step):
# pylint: disable=arguments-differ
# build resources as i currently know
container = ResourceContainer(
sdram=SDRAMResource(self.get_sdram_usage_for_atoms(vertex_slice)),
dtcm=DTCMResource(self.get_dtcm_usage_for_atoms()),
cpu_cycles=CPUCyclesPerTickResource(
self.get_cpu_usage_for_atoms()))
recording_sizes = recording_utilities.get_recording_region_sizes(
[
self._spike_recorder.get_sdram_usage_in_bytes(
vertex_slice.n_atoms,
self._max_spikes_per_ts(vertex_slice, n_machine_time_steps,
machine_time_step),
self._N_POPULATION_RECORDING_REGIONS) *
n_machine_time_steps
], self._minimum_buffer_sdram, self._maximum_sdram_for_buffering,
self._using_auto_pause_and_resume)
container.extend(
recording_utilities.get_recording_resources(
recording_sizes, self._receive_buffer_host,
self._receive_buffer_port))
return container
@property
def n_atoms(self):
return self._n_atoms
@inject_items({
"n_machine_time_steps": "TotalMachineTimeSteps",
"machine_time_step": "MachineTimeStep"
})
@overrides(
ApplicationVertex.create_machine_vertex,
additional_arguments={"n_machine_time_steps", "machine_time_step"})
def create_machine_vertex(self,
vertex_slice,
resources_required,
n_machine_time_steps,
machine_time_step,
label=None,
constraints=None):
# pylint: disable=too-many-arguments, arguments-differ
SpikeSourcePoissonVertex._n_poisson_machine_vertices += 1
buffered_sdram_per_timestep =\
self._spike_recorder.get_sdram_usage_in_bytes(
vertex_slice.n_atoms, self._max_spikes_per_ts(
vertex_slice, n_machine_time_steps, machine_time_step), 1)
minimum_buffer_sdram = recording_utilities.get_minimum_buffer_sdram(
[buffered_sdram_per_timestep * n_machine_time_steps],
self._minimum_buffer_sdram)
return SpikeSourcePoissonMachineVertex(resources_required,
self._spike_recorder.record,
minimum_buffer_sdram[0],
buffered_sdram_per_timestep,
constraints, label)
@property
def rate(self):
return self._rate
@rate.setter
def rate(self, rate):
new_rate = utility_calls.convert_param_to_numpy(rate, self._n_atoms)
self._rate_change = new_rate - self._rate
self._rate = new_rate
@property
def start(self):
return self._start
@start.setter
def start(self, start):
self._start = utility_calls.convert_param_to_numpy(
start, self._n_atoms)
@property
def duration(self):
return self._duration
@duration.setter
def duration(self, duration):
self._duration = utility_calls.convert_param_to_numpy(
duration, self._n_atoms)
@property
def seed(self):
return self._seed
@seed.setter
def seed(self, seed):
self._seed = seed
self._kiss_seed = dict()
self._rng = None
@staticmethod
def get_params_bytes(vertex_slice):
""" Gets the size of the poisson parameters in bytes
:param vertex_slice:
"""
return (PARAMS_BASE_WORDS +
(vertex_slice.n_atoms * PARAMS_WORDS_PER_NEURON)) * 4
def reserve_memory_regions(self, spec, placement, graph_mapper):
""" Reserve memory regions for poisson source parameters and output\
buffer.
:param spec: the data specification writer
:param placement: the location this vertex resides on in the machine
:param graph_mapper: the mapping between app and machine graphs
:return: None
"""
spec.comment("\nReserving memory space for data regions:\n\n")
# Reserve memory:
spec.reserve_memory_region(region=_REGIONS.SYSTEM_REGION.value,
size=SYSTEM_BYTES_REQUIREMENT,
label='setup')
# reserve poisson params dsg region
self._reserve_poisson_params_region(placement, graph_mapper, spec)
spec.reserve_memory_region(
region=_REGIONS.SPIKE_HISTORY_REGION.value,
size=recording_utilities.get_recording_header_size(1),
label="Recording")
placement.vertex.reserve_provenance_data_region(spec)
def _reserve_poisson_params_region(self, placement, graph_mapper, spec):
""" does the allocation for the poisson params region itself, as\
it can be reused for setters after an initial run
:param placement: the location on machine for this vertex
:param graph_mapper: the mapping between machine and application graphs
:param spec: the dsg writer
:return: None
"""
spec.reserve_memory_region(region=_REGIONS.POISSON_PARAMS_REGION.value,
size=self.get_params_bytes(
graph_mapper.get_slice(
placement.vertex)),
label='PoissonParams')
def _write_poisson_parameters(self, spec, graph, placement, routing_info,
vertex_slice, machine_time_step,
time_scale_factor):
""" Generate Neuron Parameter data for Poisson spike sources
:param spec: the data specification writer
:param key: the routing key for this vertex
:param vertex_slice:\
the slice of atoms a machine vertex holds from its application\
vertex
:param machine_time_step: the time between timer tick updates.
:param time_scale_factor:\
the scaling between machine time step and real time
:return: None
"""
# pylint: disable=too-many-arguments, too-many-locals
spec.comment(
"\nWriting Neuron Parameters for {} poisson sources:\n".format(
vertex_slice.n_atoms))
# Set the focus to the memory region 2 (neuron parameters):
spec.switch_write_focus(_REGIONS.POISSON_PARAMS_REGION.value)
# Write Key info for this core:
key = routing_info.get_first_key_from_pre_vertex(
placement.vertex, constants.SPIKE_PARTITION_ID)
spec.write_value(data=1 if key is not None else 0)
spec.write_value(data=key if key is not None else 0)
# Write the incoming mask if there is one
in_edges = graph.get_edges_ending_at_vertex_with_partition_name(
placement.vertex, constants.LIVE_POISSON_CONTROL_PARTITION_ID)
if len(in_edges) > 1:
raise ConfigurationException(
"Only one control edge can end at a Poisson vertex")
incoming_mask = 0
if len(in_edges) == 1:
in_edge = in_edges[0]
# Get the mask of the incoming keys
incoming_mask = \
routing_info.get_routing_info_for_edge(in_edge).first_mask
incoming_mask = ~incoming_mask & 0xFFFFFFFF
spec.write_value(incoming_mask)
# Write the random back off value
spec.write_value(
random.randint(
0,
min(self._n_poisson_machine_vertices,
MICROSECONDS_PER_SECOND // machine_time_step)))
# Write the number of microseconds between sending spikes
total_mean_rate = numpy.sum(self._rate)
if total_mean_rate > 0:
max_spikes = numpy.sum(
scipy.stats.poisson.ppf(1.0 - (1.0 / self._rate), self._rate))
spikes_per_timestep = (
max_spikes / (MICROSECONDS_PER_SECOND // machine_time_step))
# avoid a possible division by zero / small number (which may
# result in a value that doesn't fit in a uint32) by only
# setting time_between_spikes if spikes_per_timestep is > 1
time_between_spikes = 0.0
if spikes_per_timestep > 1:
time_between_spikes = (
(machine_time_step * time_scale_factor) /
(spikes_per_timestep * 2.0))
spec.write_value(data=int(time_between_spikes))
else:
# If the rate is 0 or less, set a "time between spikes" of 1
# to ensure that some time is put between spikes in event
# of a rate change later on
spec.write_value(data=1)
# Write the number of seconds per timestep (unsigned long fract)
spec.write_value(data=float(machine_time_step) /
MICROSECONDS_PER_SECOND,
data_type=DataType.U032)
# Write the number of timesteps per second (accum)
spec.write_value(data=MICROSECONDS_PER_SECOND /
float(machine_time_step),
data_type=DataType.S1615)
# Write the slow-rate-per-tick-cutoff (accum)
spec.write_value(data=SLOW_RATE_PER_TICK_CUTOFF,
data_type=DataType.S1615)
# Write the lo_atom id
spec.write_value(data=vertex_slice.lo_atom)
# Write the number of sources
spec.write_value(data=vertex_slice.n_atoms)
# Write the random seed (4 words), generated randomly!
kiss_key = (vertex_slice.lo_atom, vertex_slice.hi_atom)
if kiss_key not in self._kiss_seed:
if self._rng is None:
self._rng = numpy.random.RandomState(self._seed)
self._kiss_seed[kiss_key] = [
self._rng.randint(-0x80000000, 0x7FFFFFFF) + 0x80000000
for _ in range(4)
]
for value in self._kiss_seed[kiss_key]:
spec.write_value(data=value)
# Compute the start times in machine time steps
start = self._start[vertex_slice.as_slice]
start_scaled = self._convert_ms_to_n_timesteps(start,
machine_time_step)
# Compute the end times as start times + duration in machine time steps
# (where duration is not None)
duration = self._duration[vertex_slice.as_slice]
end_scaled = numpy.zeros(len(duration), dtype="uint32")
none_positions = numpy.isnan(duration)
positions = numpy.invert(none_positions)
end_scaled[none_positions] = 0xFFFFFFFF
end_scaled[positions] = self._convert_ms_to_n_timesteps(
start[positions] + duration[positions], machine_time_step)
# Get the rates for the atoms
rates = self._rate[vertex_slice.as_slice].astype("float")
# Compute the spikes per tick for each atom
spikes_per_tick = (
rates * (float(machine_time_step) / MICROSECONDS_PER_SECOND))
# Determine which sources are fast and which are slow
is_fast_source = spikes_per_tick > SLOW_RATE_PER_TICK_CUTOFF
# Compute the e^-(spikes_per_tick) for fast sources to allow fast
# computation of the Poisson distribution to get the number of spikes
# per timestep
exp_minus_lambda = numpy.zeros(len(spikes_per_tick), dtype="float")
exp_minus_lambda[is_fast_source] = numpy.exp(
-1.0 * spikes_per_tick[is_fast_source])
# Compute the inter-spike-interval for slow sources to get the average
# number of timesteps between spikes
isi_val = numpy.zeros(len(spikes_per_tick), dtype="float")
elements = numpy.logical_not(is_fast_source) & (spikes_per_tick > 0)
isi_val[elements] = 1.0 / spikes_per_tick[elements]
# Get the time to spike value
time_to_spike = self._time_to_spike[vertex_slice.as_slice]
changed_rates = (
self._rate_change[vertex_slice.as_slice].astype("bool") & elements)
time_to_spike[changed_rates] = 0.0
# Merge the arrays as parameters per atom
data = numpy.dstack(
(start_scaled.astype("uint32"), end_scaled.astype("uint32"),
is_fast_source.astype("uint32"),
(exp_minus_lambda * (2**32)).astype("uint32"),
(isi_val * (2**15)).astype("uint32"),
(time_to_spike * (2**15)).astype("uint32")))[0]
spec.write_array(data)
@staticmethod
def _convert_ms_to_n_timesteps(value, machine_time_step):
return numpy.round(
value * (MICROSECONDS_PER_MILLISECOND / float(machine_time_step)))
@staticmethod
def _convert_n_timesteps_to_ms(value, machine_time_step):
return (value /
(MICROSECONDS_PER_MILLISECOND / float(machine_time_step)))
@overrides(AbstractSpikeRecordable.is_recording_spikes)
def is_recording_spikes(self):
return self._spike_recorder.record
@overrides(AbstractSpikeRecordable.set_recording_spikes)
def set_recording_spikes(self,
new_state=True,
sampling_interval=None,
indexes=None):
if sampling_interval is not None:
logger.warning("Sampling interval currently not supported for "
"SpikeSourcePoisson so being ignored")
if indexes is not None:
logger.warning("indexes not supported for "
"SpikeSourcePoisson so being ignored")
self._spike_recorder.record = new_state
@overrides(AbstractSpikeRecordable.get_spikes_sampling_interval)
def get_spikes_sampling_interval(self):
return globals_variables.get_simulator().machine_time_step
def get_sdram_usage_for_atoms(self, vertex_slice):
""" calculates total sdram usage for a set of atoms
:param vertex_slice: the atoms to calculate sdram usage for
:return: sdram usage as a number of bytes
"""
poisson_params_sz = self.get_params_bytes(vertex_slice)
total_size = (
SYSTEM_BYTES_REQUIREMENT +
SpikeSourcePoissonMachineVertex.get_provenance_data_size(0) +
poisson_params_sz)
total_size += self._get_number_of_mallocs_used_by_dsg() * \
SARK_PER_MALLOC_SDRAM_USAGE
return total_size
def _get_number_of_mallocs_used_by_dsg(self):
""" Works out how many allocation requests are required by the tools
:return: the number of allocation requests
"""
standard_mallocs = self._DEFAULT_MALLOCS_USED
if self._spike_recorder.record:
standard_mallocs += 1
return standard_mallocs
@staticmethod
def get_dtcm_usage_for_atoms():
return 0
@staticmethod
def get_cpu_usage_for_atoms():
return 0
@inject_items({
"machine_time_step": "MachineTimeStep",
"time_scale_factor": "TimeScaleFactor",
"graph_mapper": "MemoryGraphMapper",
"routing_info": "MemoryRoutingInfos",
"graph": "MemoryMachineGraph"
})
@overrides(AbstractRewritesDataSpecification.regenerate_data_specification,
additional_arguments={
"machine_time_step", "time_scale_factor", "graph_mapper",
"routing_info", "graph"
})
def regenerate_data_specification(self, spec, placement, machine_time_step,
time_scale_factor, graph_mapper,
routing_info, graph):
# pylint: disable=too-many-arguments, arguments-differ
# reserve the neuron parameters data region
self._reserve_poisson_params_region(placement, graph_mapper, spec)
# allocate parameters
self._write_poisson_parameters(spec=spec,
graph=graph,
placement=placement,
routing_info=routing_info,
vertex_slice=graph_mapper.get_slice(
placement.vertex),
machine_time_step=machine_time_step,
time_scale_factor=time_scale_factor)
# end spec
spec.end_specification()
@overrides(AbstractRewritesDataSpecification.
requires_memory_regions_to_be_reloaded)
def requires_memory_regions_to_be_reloaded(self):
return self._change_requires_neuron_parameters_reload
@overrides(AbstractRewritesDataSpecification.mark_regions_reloaded)
def mark_regions_reloaded(self):
self._change_requires_neuron_parameters_reload = False
@overrides(AbstractReadParametersBeforeSet.read_parameters_from_machine)
def read_parameters_from_machine(self, transceiver, placement,
vertex_slice):
# locate sdram address to where the neuron parameters are stored
poisson_parameter_region_sdram_address = \
helpful_functions.locate_memory_region_for_placement(
placement, _REGIONS.POISSON_PARAMS_REGION.value, transceiver)
# shift past the extra stuff before neuron parameters that we don't
# need to read
poisson_parameter_parameters_sdram_address = \
poisson_parameter_region_sdram_address + \
START_OF_POISSON_GENERATOR_PARAMETERS
# get size of poisson params
size_of_region = self.get_params_bytes(vertex_slice)
size_of_region -= START_OF_POISSON_GENERATOR_PARAMETERS
# get data from the machine
byte_array = transceiver.read_memory(
placement.x, placement.y,
poisson_parameter_parameters_sdram_address, size_of_region)
# Convert the data to parameter values
(start, end, is_fast_source, exp_minus_lambda, isi,
time_to_next_spike) = _PoissonStruct.read_data(
byte_array, 0, vertex_slice.n_atoms)
# Convert start values as timesteps into milliseconds
self._start[vertex_slice.as_slice] = self._convert_n_timesteps_to_ms(
start, self._machine_time_step)
# Convert end values as timesteps to durations in milliseconds
self._duration[vertex_slice.as_slice] = (
self._convert_n_timesteps_to_ms(end, self._machine_time_step) -
self._start[vertex_slice.as_slice])
# Work out the spikes per tick depending on if the source is slow
# or fast
is_fast_source = is_fast_source == 1.0
spikes_per_tick = numpy.zeros(len(is_fast_source), dtype="float")
spikes_per_tick[is_fast_source] = numpy.log(
exp_minus_lambda[is_fast_source]) * -1.0
slow_elements = isi > 0
spikes_per_tick[slow_elements] = 1.0 / isi[slow_elements]
# Convert spikes per tick to rates
self._rate[vertex_slice.as_slice] = (
spikes_per_tick *
(MICROSECONDS_PER_SECOND / float(self._machine_time_step)))
# Store the updated time until next spike so that it can be
# rewritten when the parameters are loaded
self._time_to_spike[vertex_slice.as_slice] = time_to_next_spike
@inject_items({
"machine_time_step": "MachineTimeStep",
"time_scale_factor": "TimeScaleFactor",
"graph_mapper": "MemoryGraphMapper",
"routing_info": "MemoryRoutingInfos",
"tags": "MemoryTags",
"n_machine_time_steps": "TotalMachineTimeSteps",
"graph": "MemoryMachineGraph"
})
@overrides(AbstractGeneratesDataSpecification.generate_data_specification,
additional_arguments={
"machine_time_step", "time_scale_factor", "graph_mapper",
"routing_info", "tags", "n_machine_time_steps", "graph"
})
def generate_data_specification(self, spec, placement, machine_time_step,
time_scale_factor, graph_mapper,
routing_info, tags, n_machine_time_steps,
graph):
# pylint: disable=too-many-arguments, arguments-differ
self._machine_time_step = machine_time_step
vertex = placement.vertex
vertex_slice = graph_mapper.get_slice(vertex)
spec.comment("\n*** Spec for SpikeSourcePoisson Instance ***\n\n")
# Reserve SDRAM space for memory areas:
self.reserve_memory_regions(spec, placement, graph_mapper)
# write setup data
spec.switch_write_focus(_REGIONS.SYSTEM_REGION.value)
spec.write_array(
simulation_utilities.get_simulation_header_array(
self.get_binary_file_name(), machine_time_step,
time_scale_factor))
# write recording data
ip_tags = tags.get_ip_tags_for_vertex(vertex)
spec.switch_write_focus(_REGIONS.SPIKE_HISTORY_REGION.value)
recorded_region_sizes = recording_utilities.get_recorded_region_sizes([
self._spike_recorder.get_sdram_usage_in_bytes(
vertex_slice.n_atoms,
self._max_spikes_per_ts(vertex_slice, n_machine_time_steps,
machine_time_step),
n_machine_time_steps)
], self._maximum_sdram_for_buffering)
spec.write_array(
recording_utilities.get_recording_header_array(
recorded_region_sizes, self._time_between_requests,
self._buffer_size_before_receive, ip_tags))
# write parameters
self._write_poisson_parameters(spec, graph, placement, routing_info,
vertex_slice, machine_time_step,
time_scale_factor)
# End-of-Spec:
spec.end_specification()
@overrides(AbstractHasAssociatedBinary.get_binary_file_name)
def get_binary_file_name(self):
return "spike_source_poisson.aplx"
@overrides(AbstractHasAssociatedBinary.get_binary_start_type)
def get_binary_start_type(self):
return ExecutableType.USES_SIMULATION_INTERFACE
@overrides(AbstractSpikeRecordable.get_spikes)
def get_spikes(self, placements, graph_mapper, buffer_manager,
machine_time_step):
return self._spike_recorder.get_spikes(
self.label, buffer_manager,
SpikeSourcePoissonVertex.SPIKE_RECORDING_REGION_ID, placements,
graph_mapper, self, machine_time_step)
@overrides(AbstractProvidesOutgoingPartitionConstraints.
get_outgoing_partition_constraints)
def get_outgoing_partition_constraints(self, partition):
return [ContiguousKeyRangeContraint()]
@overrides(AbstractSpikeRecordable.clear_spike_recording)
def clear_spike_recording(self, buffer_manager, placements, graph_mapper):
machine_vertices = graph_mapper.get_machine_vertices(self)
for machine_vertex in machine_vertices:
placement = placements.get_placement_of_vertex(machine_vertex)
buffer_manager.clear_recorded_data(
placement.x, placement.y, placement.p,
SpikeSourcePoissonVertex.SPIKE_RECORDING_REGION_ID)
def describe(self):
"""
Returns a human-readable description of the cell or synapse type.
The output may be customised by specifying a different template
together with an associated template engine
(see ``pyNN.descriptions``).
If template is None, then a dictionary containing the template context
will be returned.
"""
parameters = dict()
for parameter_name in self._model.default_parameters:
parameters[parameter_name] = self.get_value(parameter_name)
context = {
"name": self._model_name,
"default_parameters": self._model.default_parameters,
"default_initial_values": self._model.default_parameters,
"parameters": parameters,
}
return context
class SpikeSourcePoissonVertex(
TDMAAwareApplicationVertex, AbstractSpikeRecordable,
AbstractProvidesOutgoingPartitionConstraints,
AbstractChangableAfterRun, SimplePopulationSettable,
ProvidesKeyToAtomMappingImpl, LegacyPartitionerAPI):
""" A Poisson Spike source object
"""
__slots__ = [
"__change_requires_mapping",
"__duration",
"__model",
"__model_name",
"__n_atoms",
"__rate",
"__rng",
"__seed",
"__spike_recorder",
"__start",
"__time_to_spike",
"__kiss_seed", # dict indexed by vertex slice
"__n_subvertices",
"__n_data_specs",
"__max_rate",
"__rate_change",
"__n_profile_samples",
"__data",
"__is_variable_rate",
"__max_spikes",
"__outgoing_projections"
]
SPIKE_RECORDING_REGION_ID = 0
def __init__(self,
n_neurons,
constraints,
label,
seed,
max_atoms_per_core,
model,
rate=None,
start=None,
duration=None,
rates=None,
starts=None,
durations=None,
max_rate=None,
splitter=None):
"""
:param int n_neurons:
:param constraints:
:type constraints:
iterable(~pacman.model.constraints.AbstractConstraint)
:param str label:
:param float seed:
:param int max_atoms_per_core:
:param ~spynnaker.pyNN.models.spike_source.SpikeSourcePoisson model:
:param iterable(float) rate:
:param iterable(int) start:
:param iterable(int) duration:
:param splitter:
:type splitter:
~pacman.model.partitioner_splitters.abstract_splitters.AbstractSplitterCommon
"""
# pylint: disable=too-many-arguments
super().__init__(label, constraints, max_atoms_per_core, splitter)
# atoms params
self.__n_atoms = self.round_n_atoms(n_neurons, "n_neurons")
self.__model_name = "SpikeSourcePoisson"
self.__model = model
self.__seed = seed
self.__kiss_seed = dict()
self.__n_subvertices = 0
self.__n_data_specs = 0
# check for changes parameters
self.__change_requires_mapping = True
self.__spike_recorder = MultiSpikeRecorder()
# Check for disallowed pairs of parameters
if (rates is not None) and (rate is not None):
raise Exception("Exactly one of rate and rates can be specified")
if (starts is not None) and (start is not None):
raise Exception("Exactly one of start and starts can be specified")
if (durations is not None) and (duration is not None):
raise Exception(
"Exactly one of duration and durations can be specified")
if rate is None and rates is None:
raise Exception("One of rate or rates must be specified")
# Normalise the parameters
self.__is_variable_rate = rates is not None
if rates is None:
if hasattr(rate, "__len__"):
# Single rate per neuron for whole simulation
rates = [numpy.array([r]) for r in rate]
else:
# Single rate for all neurons for whole simulation
rates = numpy.array([rate])
elif hasattr(rates[0], "__len__"):
# Convert each list to numpy array
rates = [numpy.array(r) for r in rates]
else:
rates = numpy.array(rates)
if starts is None and start is not None:
if hasattr(start, "__len__"):
starts = [numpy.array([s]) for s in start]
elif start is None:
starts = numpy.array([0])
else:
starts = numpy.array([start])
elif starts is not None and hasattr(starts[0], "__len__"):
starts = [numpy.array(s) for s in starts]
elif starts is not None:
starts = numpy.array(starts)
if durations is None and duration is not None:
if hasattr(duration, "__len__"):
durations = [numpy.array([d]) for d in duration]
else:
durations = numpy.array([duration])
elif durations is not None and hasattr(durations[0], "__len__"):
durations = [numpy.array(d) for d in durations]
elif durations is not None:
durations = numpy.array(durations)
else:
if hasattr(rates[0], "__len__"):
durations = [
numpy.array([None for r in _rate]) for _rate in rates
]
else:
durations = numpy.array([None for _rate in rates])
# Check that there is either one list for all neurons,
# or one per neuron
if hasattr(rates[0], "__len__") and len(rates) != n_neurons:
raise Exception(
"Must specify one rate for all neurons or one per neuron")
if (starts is not None and hasattr(starts[0], "__len__")
and len(starts) != n_neurons):
raise Exception(
"Must specify one start for all neurons or one per neuron")
if (durations is not None and hasattr(durations[0], "__len__")
and len(durations) != n_neurons):
raise Exception(
"Must specify one duration for all neurons or one per neuron")
# Check that for each rate there is a start and duration if needed
# TODO: Could be more efficient for case where parameters are not one
# per neuron
for i in range(n_neurons):
rate_set = rates
if hasattr(rates[0], "__len__"):
rate_set = rates[i]
if not hasattr(rate_set, "__len__"):
raise Exception("Multiple rates must be a list")
if starts is None and len(rate_set) > 1:
raise Exception("When multiple rates are specified,"
" each must have a start")
elif starts is not None:
start_set = starts
if hasattr(starts[0], "__len__"):
start_set = starts[i]
if len(start_set) != len(rate_set):
raise Exception("Each rate must have a start")
if any(s is None for s in start_set):
raise Exception("Start must not be None")
if durations is not None:
duration_set = durations
if hasattr(durations[0], "__len__"):
duration_set = durations[i]
if len(duration_set) != len(rate_set):
raise Exception("Each rate must have its own duration")
if hasattr(rates[0], "__len__"):
time_to_spike = [
numpy.array([0 for _ in range(len(rates[i]))])
for i in range(len(rates))
]
else:
time_to_spike = numpy.array([0 for _ in range(len(rates))])
self.__data = SpynnakerRangeDictionary(n_neurons)
self.__data["rates"] = SpynnakerRangedList(
n_neurons,
rates,
use_list_as_value=not hasattr(rates[0], "__len__"))
self.__data["starts"] = SpynnakerRangedList(
n_neurons,
starts,
use_list_as_value=not hasattr(starts[0], "__len__"))
self.__data["durations"] = SpynnakerRangedList(
n_neurons,
durations,
use_list_as_value=not hasattr(durations[0], "__len__"))
self.__data["time_to_spike"] = SpynnakerRangedList(
n_neurons,
time_to_spike,
use_list_as_value=not hasattr(time_to_spike[0], "__len__"))
self.__rng = numpy.random.RandomState(seed)
self.__rate_change = numpy.zeros(n_neurons)
self.__n_profile_samples = get_config_int("Reports",
"n_profile_samples")
# Prepare for recording, and to get spikes
self.__spike_recorder = MultiSpikeRecorder()
all_rates = list(_flatten(self.__data["rates"]))
self.__max_rate = max_rate
if max_rate is None and len(all_rates):
self.__max_rate = numpy.amax(all_rates)
elif max_rate is None:
self.__max_rate = 0
total_rate = numpy.sum(all_rates)
self.__max_spikes = 0
if total_rate > 0:
# Note we have to do this per rate, as the whole array is not numpy
max_rates = numpy.array(
[numpy.max(r) for r in self.__data["rates"]])
self.__max_spikes = numpy.sum(
scipy.stats.poisson.ppf(1.0 - (1.0 / max_rates), max_rates))
# Keep track of how many outgoing projections exist
self.__outgoing_projections = list()
def add_outgoing_projection(self, projection):
""" Add an outgoing projection from this vertex
:param PyNNProjectionCommon projection: The projection to add
"""
self.__outgoing_projections.append(projection)
@property
def outgoing_projections(self):
""" The projections outgoing from this vertex
:rtype: list(PyNNProjectionCommon)
"""
return self.__outgoing_projections
@property
def n_profile_samples(self):
return self.__n_profile_samples
@property
def rate(self):
if self.__is_variable_rate:
raise Exception("Get variable rate poisson rates with .rates")
return list(_flatten(self.__data["rates"]))
@rate.setter
def rate(self, rate):
if self.__is_variable_rate:
raise Exception("Cannot set rate of a variable rate poisson")
self.__rate_change = rate - numpy.array(
list(_flatten(self.__data["rates"])))
# Normalise parameter
if hasattr(rate, "__len__"):
# Single rate per neuron for whole simulation
self.__data["rates"].set_value([numpy.array([r]) for r in rate])
else:
# Single rate for all neurons for whole simulation
self.__data["rates"].set_value(numpy.array([rate]),
use_list_as_value=True)
all_rates = list(_flatten(self.__data["rates"]))
new_max = 0
if len(all_rates):
new_max = numpy.amax(all_rates)
if self.__max_rate is None:
self.__max_rate = new_max
# Setting record forces reset so OK to go over if not recording
elif self.__spike_recorder.record and new_max > self.__max_rate:
logger.info('Increasing spike rate while recording requires a '
'"reset unless additional_parameters "max_rate" is '
'set')
self.__change_requires_mapping = True
self.__max_rate = new_max
@property
def start(self):
return self.__data["starts"]
@start.setter
def start(self, start):
if self.__is_variable_rate:
raise Exception("Cannot set start of a variable rate poisson")
# Normalise parameter
if hasattr(start, "__len__"):
# Single start per neuron for whole simulation
self.__data["starts"].set_value([numpy.array([s]) for s in start])
else:
# Single start for all neurons for whole simulation
self.__data["starts"].set_value(numpy.array([start]),
use_list_as_value=True)
@property
def duration(self):
return self.__data["durations"]
@duration.setter
def duration(self, duration):
if self.__is_variable_rate:
raise Exception("Cannot set duration of a variable rate poisson")
# Normalise parameter
if hasattr(duration, "__len__"):
# Single duration per neuron for whole simulation
self.__data["durations"].set_value(
[numpy.array([d]) for d in duration])
else:
# Single duration for all neurons for whole simulation
self.__data["durations"].set_value(numpy.array([duration]),
use_list_as_value=True)
@property
def rates(self):
return self.__data["rates"]
@rates.setter
def rates(self, _rates):
if self.__is_variable_rate:
raise Exception("Cannot set rates of a variable rate poisson")
raise Exception("Set the rate of a Poisson source using rate")
@property
def starts(self):
return self.__data["starts"]
@starts.setter
def starts(self, _starts):
if self.__is_variable_rate:
raise Exception("Cannot set starts of a variable rate poisson")
raise Exception("Set the start of a Poisson source using start")
@property
def durations(self):
return self.__data["durations"]
@durations.setter
def durations(self, _durations):
if self.__is_variable_rate:
raise Exception("Cannot set durations of a variable rate poisson")
raise Exception("Set the duration of a Poisson source using duration")
@property
def time_to_spike(self):
return self.__data["time_to_spike"]
@property
def rate_change(self):
return self.__rate_change
@property
@overrides(AbstractChangableAfterRun.requires_mapping)
def requires_mapping(self):
return self.__change_requires_mapping
@overrides(AbstractChangableAfterRun.mark_no_changes)
def mark_no_changes(self):
self.__change_requires_mapping = False
@overrides(SimplePopulationSettable.set_value)
def set_value(self, key, value):
super().set_value(key, value)
for machine_vertex in self.machine_vertices:
if isinstance(machine_vertex, AbstractRewritesDataSpecification):
machine_vertex.set_reload_required(True)
def max_spikes_per_ts(self):
ts_per_second = (MICRO_TO_SECOND_CONVERSION / machine_time_step())
if float(self.__max_rate) / ts_per_second < \
SLOW_RATE_PER_TICK_CUTOFF:
return 1
# Experiments show at 1000 this result is typically higher than actual
chance_ts = 1000
max_spikes_per_ts = scipy.stats.poisson.ppf(
1.0 - (1.0 / float(chance_ts)),
float(self.__max_rate) / ts_per_second)
return int(math.ceil(max_spikes_per_ts)) + 1.0
def get_recording_sdram_usage(self, vertex_slice):
"""
:param ~pacman.model.graphs.common.Slice vertex_slice:
"""
variable_sdram = self.__spike_recorder.get_sdram_usage_in_bytes(
vertex_slice.n_atoms, self.max_spikes_per_ts())
constant_sdram = ConstantSDRAM(variable_sdram.per_timestep *
OVERFLOW_TIMESTEPS_FOR_SDRAM)
return variable_sdram + constant_sdram
@overrides(LegacyPartitionerAPI.get_resources_used_by_atoms)
def get_resources_used_by_atoms(self, vertex_slice):
"""
:param ~pacman.model.graphs.common.Slice vertex_slice:
"""
# pylint: disable=arguments-differ
poisson_params_sz = get_rates_bytes(vertex_slice, self.__data["rates"])
sdram_sz = get_sdram_edge_params_bytes(vertex_slice)
other = ConstantSDRAM(
SYSTEM_BYTES_REQUIREMENT +
SpikeSourcePoissonMachineVertex.get_provenance_data_size(0) +
poisson_params_sz + self.tdma_sdram_size_in_bytes +
recording_utilities.get_recording_header_size(1) +
recording_utilities.get_recording_data_constant_size(1) +
profile_utils.get_profile_region_size(self.__n_profile_samples) +
sdram_sz)
recording = self.get_recording_sdram_usage(vertex_slice)
# build resources as i currently know
container = ResourceContainer(sdram=recording + other,
dtcm=DTCMResource(
self.get_dtcm_usage_for_atoms()),
cpu_cycles=CPUCyclesPerTickResource(
self.get_cpu_usage_for_atoms()))
return container
@property
def n_atoms(self):
return self.__n_atoms
@overrides(LegacyPartitionerAPI.create_machine_vertex)
def create_machine_vertex(self,
vertex_slice,
resources_required,
label=None,
constraints=None):
# pylint: disable=too-many-arguments, arguments-differ
index = self.__n_subvertices
self.__n_subvertices += 1
return SpikeSourcePoissonMachineVertex(resources_required,
self.__spike_recorder.record,
constraints, label, self,
vertex_slice, index)
@property
def max_rate(self):
return self.__max_rate
@property
def seed(self):
return self.__seed
@seed.setter
def seed(self, seed):
self.__seed = seed
self.__kiss_seed = dict()
self.__rng = numpy.random.RandomState(seed)
@overrides(AbstractSpikeRecordable.is_recording_spikes)
def is_recording_spikes(self):
return self.__spike_recorder.record
@overrides(AbstractSpikeRecordable.set_recording_spikes)
def set_recording_spikes(self,
new_state=True,
sampling_interval=None,
indexes=None):
if sampling_interval is not None:
logger.warning("Sampling interval currently not supported for "
"SpikeSourcePoisson so being ignored")
if indexes is not None:
logger.warning("indexes not supported for "
"SpikeSourcePoisson so being ignored")
if new_state and not self.__spike_recorder.record:
self.__change_requires_mapping = True
self.__spike_recorder.record = new_state
@overrides(AbstractSpikeRecordable.get_spikes_sampling_interval)
def get_spikes_sampling_interval(self):
return machine_time_step()
@staticmethod
def get_dtcm_usage_for_atoms():
return 0
@staticmethod
def get_cpu_usage_for_atoms():
return 0
def kiss_seed(self, vertex_slice):
if vertex_slice not in self.__kiss_seed:
self.__kiss_seed[vertex_slice] = create_mars_kiss_seeds(self.__rng)
return self.__kiss_seed[vertex_slice]
def update_kiss_seed(self, vertex_slice, seed):
""" updates a kiss seed from the machine
:param vertex_slice: the vertex slice to update seed of
:param seed: the seed
:rtype: None
"""
self.__kiss_seed[vertex_slice] = seed
@overrides(AbstractSpikeRecordable.get_spikes)
def get_spikes(self, placements, buffer_manager):
return self.__spike_recorder.get_spikes(
self.label, buffer_manager,
SpikeSourcePoissonVertex.SPIKE_RECORDING_REGION_ID, placements,
self)
@overrides(AbstractProvidesOutgoingPartitionConstraints.
get_outgoing_partition_constraints)
def get_outgoing_partition_constraints(self, partition):
return [ContiguousKeyRangeContraint()]
@overrides(AbstractSpikeRecordable.clear_spike_recording)
def clear_spike_recording(self, buffer_manager, placements):
for machine_vertex in self.machine_vertices:
placement = placements.get_placement_of_vertex(machine_vertex)
buffer_manager.clear_recorded_data(
placement.x, placement.y, placement.p,
SpikeSourcePoissonVertex.SPIKE_RECORDING_REGION_ID)
def describe(self):
""" Return a human-readable description of the cell or synapse type.
The output may be customised by specifying a different template\
together with an associated template engine\
(see :py:mod:`pyNN.descriptions`).
If template is None, then a dictionary containing the template context\
will be returned.
:rtype: dict(str, ...)
"""
parameters = dict()
for parameter_name in self.__model.default_parameters:
parameters[parameter_name] = self.get_value(parameter_name)
context = {
"name": self.__model_name,
"default_parameters": self.__model.default_parameters,
"default_initial_values": self.__model.default_parameters,
"parameters": parameters,
}
return context
@overrides(TDMAAwareApplicationVertex.get_n_cores)
def get_n_cores(self):
return len(self._splitter.get_out_going_slices()[0])